humans&
3 Intel
humans& is a frontier AI research and product lab building human-centric AI: systems meant to support human collaboration and coordination rather than replace people. It was founded in late 2025 by researchers from Anthropic, xAI and Google together with Stanford professor Noah Goodman; co-founders include Andi Peng, Eric Zelikman, Yuchen He and early Google employee Georges Harik. The company announced itself publicly in January 2026 alongside a large seed round.
Is this you? Sign in with X to claim this profile.
Intel
humans& released Persimmon, a research-preview model post-trained from NVIDIA's 550B Nemotron 3 Ultra specifically to simulate realistic human conversation, built because AI judges could reliably flag assistant-simulated dialogue as artificial.
humans& releases Ampersand v0.1 via its agent API, a model trained on simulated multi-month interactions with a learned user model to handle group collaboration. The announcement claims gains over Fable and other frontier models in its own evals; the linked blog and docs were not reviewed.
humans& open-sourced a hardware-native 4-bit reinforcement learning recipe that fixes instabilities from policy error and gradient mismatch, developed with NVIDIA and RadixArk to speed up long-horizon multi-agent RL training.